By Mak Nurkić Kačapor, Junior at British International School, Washington DC and co-creator of DUXO Project
and Justin Yang, Junior at Walt Whitman High School, Bethesda, Maryland and co-creator of DUXO Project
THE PROJECT AND MISSION
The DUXO Project was founded by a group of students from various backgrounds with a shared passion for demining advocacy. With both personal connections to the issue and the resources, location, and network to create tangible change, we devised a mission:
To educate the public on the afflictions caused by landmines to a large part of human society, and to design an efficient, accurate, and robust drone solution to landmine detection, in order to increase efficiency, effectiveness, and ensure safety in the demining industry.
The first half of our mission statement speaks to the notable lack of awareness and involvement by the average person in the landmine issue in the past decades. Whilst more than 60 countries and territories continue to face the challenges of landmine contamination and clearance, across the world – particularly in countries where social media and interest aggregation are at their most powerful – the general public remains uninformed and uninvolved in mine action.
Although international NGOs and state bodies have effectively established both a domestic and international-level structure for mine action, the goal of a landmine-free world is many decades away. So, what can public involvement do about this?
Two of the largest obstacles facing mine action are policies and funding. While mine clearance continues in many countries around the world, the stockpiling, production, and in some cases use of mines and Improvised Explosive Devices (IEDs) continue to take place. Additionally, the same military and defense budgets that fund these counterproductive activities dwarf the international funding for disarmament, including mine action.
As the world bore witness with various social issues across the world in the past decade, social media and the internet have completely changed the way that the public gains access to information and calls for change; to push for changes in policy and increase funding for a cause, the public needs to understand and support it. We recognize this very advantage in our generation and the power that social media and the internet have for the rapid dissemination of ideas and influence, amongst people of many ages, all of whom now have the previously unimaginable ability to create change. To bring the issue of landmines back to these social discussions, we’ve designed interactive infographics and publicly accessible datasets on our website, duxo.org, that implement data pulled from databases and reports across the mine action industry into modern, engaging tools for the general public. Our next milestone for this action is gaining direct access to the most accurate and up-to-date survey information for each landmine-affected region, which would improve both the quality and quantity of the information we share.
However, in addition to the distribution of key mine action-related information and advocacy, we have also been developing a technical solution to the detection process in landmine clearance. Over the past years, we have researched and developed our patent-pending drone technology for detecting subsurface structures, with a particular emphasis on landmines, implementing three detection methods into one cohesive system. Ground-Penetrating Radar, in conjunction with Metal Detection and Infrared Imaging, are all carried on a single aerial platform where each output is considered by a central network and a landmine likelihood score is established. This is the second part of our mission, where advocacy turns to tangible change.
Systems and method to detect presence of buried landmines in a suspect area: A drone is outfitted with a Ground-Penetrating Radar unit, an infrared camera, and a metal detector mounted onto a leveling platform. The drone is flown over the suspect area while maintaining the leveling platform horizontal to the detection surface. Signals from the ground penetrating radar, infrared camera, and metal detector are converted into radargrams, thermal images, and metal grams. Convolutional Neural Networks are applied to each of the produced images to detect anomalies. These anomalies are identified on a probability scale for landmines.
Ground Penetrating Radar (GPR) is a non-intrusive detection method for structures underneath the surface. GPR uses radio wave pulses (in the spectrum from 10 MHz to 2.6 GHz) to penetrate a large range of surfaces, from sand, rock, soil, ice. These pulses reflect, refract, or scatter from subsurface objects with different permittivities (from other objects and the ground). GPR will identify underground anomalies in the forms and dimensions of landmines, regardless of metallic or plastic composition.
Metal Detection in the payload will operate using either Very Low Frequency (VLF) or Pulse Induction sensing. A Pulse Induction metal detector uses a single coil to emit and receive rapid high-voltage pulses in multiple frequencies at a time. The pulses decay as they re- enter the circuit through the coil due to the resistance of the system. Metal objects, like landmines, can induce a magnetic field caused by these pulses, which decay differently than the returning pulses. A VLF metal detector uses two overlapping coils: an emitting coil to produce a magnetic field and a receiving coil to observe changes to the magnetic flux introduced by metal objects in the environment, such as landmines and landmine components. On board processing will graph and map the detected metals on the flight grid as they are located. Once the grid is completed, the drone will process the size and composition of the detected objects in a Convolutional Neural Network, which will assign a probability value for landmines based on determinants.
Infrared Thermography detects electromagnetic radiation in the infrared spectrum (300 GHz-430THz) emitted from objects, and outputs heat-based images of a subject through a lens, like a regular camera. The IR sensor in the payload will take advantage of the differences in thermophysical properties of landmines and the ground media they lie in. This applies to the detection of “Minimal Metal Mines” (MMMs), which have plastic casings that are hard to detect with metal detection. In different conditions such as weather, climate, and time of day, plastic landmines will retain and radiate different values of heat than a given soil, sand, gravel, or other ground medium around them. Because of this, when captured with IR Thermography, the surface heat levels above and around landmines are different from the ground around them. This produces anomalies on the heat map.
Mounting and Stabilization of Sensors
Currently Ground Penetrating Radar and Metal Detectors are unable to function stand-off. This is because the mix of differing frequencies are incompatible with current technology. Instead of creating a GPR and MD apparatus that can function with mixed signals, we stabilize the entire mount by suspending it from the drone with either wire or servo arms. These will be controlled by a separate gyroscope located in the mounting platform. Whenever turbulence occurs, the platform will be reoriented to stay level to the ground to ensure proper detection. An additional sensor will measure the height of the ground, which will communicate with the main propulsion system to keep the drone at a consistent height.
CNNs and Decision System
A Convolutional Neural Network (CNN) is a deep learning algorithm which takes input images, assigns importance to them using biases to determine importance of different objects and inputs, that can be differentiated from each other. This landmine detector uses three different technologies as sources of input: infrared imaging, ground-penetrating radar, and metal detection. The algorithm will be coded and designed by teaching it biases and weightages of different objects based on characteristics of landmines. The input is an image. These images are a matrix of pixel values. Here CNNs are able to apply filters that are taught in the code to capture the spatial and temporal tendencies, which is challenging with other technologies as there are too many pixels in many colors. CNNs are able to understand the image better. In the Convolutional layer, the first part of the operation is called the Kernel/filter (K), which is a smoothing method for input. It performs matrix multiplication operation between K and the image. Landmines have multiple channels that they will be detected by in terms of colors and image, in which case K will have the same depth as the image. Each of the three technologies will have their own processing through CNN as they have different inputs. Next, in the pooling layer, the objective is to size down and reduce computational power for the input. This technology will use Maximum Pooling by taking the maximum from all values that the K covers. It discards noisy activations and performs dimension-reduction smoothly and accurately. Lastly, is the Classification, the Fully Connected Layer. Once the image is in suitable form for the multi-level perceptron, the image from each of the technologies shall be flattened into a column vector. This will be fed to a feed-forward neural network and backpropagation. In the end, the inputs will be classified using a SoftMax classification technique. These classifications of input will be run and assigned a probability value as to if there is a landmine in the area covered K, provided by the exterior technology.
The system also considers “particular considerations” in a decision tree, that account for types of landmines and special situations in the field.
- Whenever (or only) GPR detects an anomaly sizable enough to be a potential mine, it is flagged and given the minimum probability rating. This is to avoid missing landmines if either Metal Detection or IR fails (false negatives).
- Whenever GPR and either other sensor detects anomalies at the same location, it is flagged and given at least a minimum landmine probability rating. This is to avoid missing landmines if either Metal Detection or IR fails (false negatives).
- If only metal detection and GPR detect anomalies, Infrared score is disregarded (metal landmine).
- If Infrared and GPR ratings are above the landmine threshold and metal is not detected, metal detection score is disregarded.
- If Infrared and GPR ratings above threshold and metal is low but not negligible, score is increased (minimal metal mine)
- If either Infrared or Metal Detection detect anomalies where GPR does not, the system is paused and reviewed for false negatives.
- If both Infrared and Metal Detection detect anomalies above the threshold where GPR does not, the system identifies a detection error, and the field survey is stopped.
A final score is given, and the value decides the landmine likelihood rating. This rating will be actively updated using Bayesian Statistical Modeling as adequate data is accumulated. We anticipate a new posterior model after every dispatch of the drone, where afterward the likelihood data gathered from alternate sources can be applied to the prior model to improve discernment. This new model is created by applying data gathered from the previous dispatch to the existing prior model using Bayesian statistics. Based on the probability score, demining teams can prioritize and coordinate safe and efficient demining operations in the field down to maneuvering around each individual piece of ordnance. ■
Please reach out if you have any questions or are interested in collaborating with any aspect of the DUXO Project. We are always open to learning more from experts in the field, and answer any questions about our work.
ABOUT THE AUTHORS